Skip to content

Commit

Permalink
better integer support
Browse files Browse the repository at this point in the history
  • Loading branch information
@chriselrod
chriselrod committed Sep 30, 2022
1 parent e838c43 commit 4a5c5ac
Show file tree
Hide file tree
Showing 3 changed files with 46 additions and 22 deletions.
44 changes: 25 additions & 19 deletions src/SLEEFPirates.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -163,8 +163,8 @@ for func in (:sin, :cos, :tan, :asin, :acos, :atan, :sinh, :cosh, :tanh,
:sin_fast, :cos_fast, :tan_fast, :asin_fast, :acos_fast, :atan_fast,# :atan2_fast,
:log_fast, :log2_fast, :log10_fast, :cbrt_fast)#, :exp, :exp2, :exp10
@eval begin
$func(a::Float16) = Float16.($func(Float32(a)))
$func(x::Real) = $func(float(x))
@inline $func(a::Float16) = Float16.($func(Float32(a)))
@inline $func(x::Real) = $func(float(x))
@inline $func(v::AbstractSIMD{W,I}) where {W,I<:Integer} = $func(float(v))
@inline $func(i::MM) = $func(Vec(i))
@inline $func(v::VecUnroll{N,1,T,T}) where {N,T} = to_vecunrollscalar($func(VectorizationBase.transpose_vecunroll(v)), StaticInt{N}())
Expand Down Expand Up @@ -205,22 +205,22 @@ end
@inline Base.FastMath.atan_fast(a::T, b::T) where {T<:AbstractSIMD} = atan_fast(a, b)
@inline Base.FastMath.atan_fast(a::AbstractSIMD, b::AbstractSIMD) = ((c,d) = promote(a,b); atan_fast(c, d))
for func in (:atan, :hypot, :pow)
func2 = func === :pow ? :^ : func
ptyp = func === :pow ? :FloatingTypes : :NativeTypes
@eval begin
$func(y::Real, x::Real) = $func(promote(float(y), float(x))...)
$func(a::Float16, b::Float16) = Float16($func(Float32(a), Float32(b)))
# @inline Base.$func2(x::AbstractSIMD{W,T}, y::Vec{W,T}) where {W,T<:Union{Float32,Float64}} = $func(x, Vec(y))
# @inline Base.$func2(x::Vec{W,T}, y::AbstractSIMD{W,T}) where {W,T<:Union{Float32,Float64}} = $func(Vec(x), y)
@inline Base.$func2(x::AbstractSIMD{W,T}, y::T) where {W,T<:Union{Float32,Float64}} = $func(x, convert(Vec{W,T}, y))
@inline Base.$func2(x::T, y::AbstractSIMD{W,T}) where {W,T<:Union{Float32,Float64}} = $func(convert(Vec{W,T}, x), y)
@inline Base.$func2(x::AbstractSIMD{W,T1}, y::T2) where {W,T1<:Union{Float32,Float64},T2<:$ptyp} = $func(x, convert(Vec{W,T1}, y))
@inline Base.$func2(x::T2, y::AbstractSIMD{W,T1}) where {W,T1<:Union{Float32,Float64},T2<:NativeTypes} = $func(convert(Vec{W,T1}, x), y)
@inline Base.$func2(x::AbstractSIMD{W,T}, y::AbstractSIMD{W,T}) where {W,T<:Union{Float32,Float64}} = $func(x, y)
@inline $func(v1::AbstractSIMD{W,I}, v2::AbstractSIMD{W,I}) where {W,I<:Integer} = $func(float(v1), float(v2))
end
func2 = func === :pow ? :^ : func
ptyp = func === :pow ? :FloatingTypes : :NativeTypes
@eval begin
@inline $func(y::Real, x::Real) = $func(promote(float(y), float(x))...)
@inline $func(a::Float16, b::Float16) = Float16($func(Float32(a), Float32(b)))
# @inline Base.$func2(x::AbstractSIMD{W,T}, y::Vec{W,T}) where {W,T<:Union{Float32,Float64}} = $func(x, Vec(y))
# @inline Base.$func2(x::Vec{W,T}, y::AbstractSIMD{W,T}) where {W,T<:Union{Float32,Float64}} = $func(Vec(x), y)
@inline Base.$func2(x::AbstractSIMD{W,T}, y::T) where {W,T<:Union{Float32,Float64}} = $func(x, convert(Vec{W,T}, y))
@inline Base.$func2(x::T, y::AbstractSIMD{W,T}) where {W,T<:Union{Float32,Float64}} = $func(convert(Vec{W,T}, x), y)
@inline Base.$func2(x::AbstractSIMD{W,T1}, y::T2) where {W,T1<:Union{Float32,Float64},T2<:$ptyp} = $func(x, convert(Vec{W,T1}, y))
@inline Base.$func2(x::T2, y::AbstractSIMD{W,T1}) where {W,T1<:Union{Float32,Float64},T2<:NativeTypes} = $func(convert(Vec{W,T1}, x), y)
@inline Base.$func2(x::AbstractSIMD{W,T}, y::AbstractSIMD{W,T}) where {W,T<:Union{Float32,Float64}} = $func(x, y)
@inline $func(v1::AbstractSIMD{W,I}, v2::AbstractSIMD{W,I}) where {W,I<:Integer} = $func(float(v1), float(v2))
end
end
ldexp(x::Float16, q::Int) = Float16(ldexpk(Float32(x), q))
@inline ldexp(x::Float16, q::Int) = Float16(ldexpk(Float32(x), q))

# @inline logit(x) = log(Base.FastMath.div_fast(x,Base.FastMath.sub_fast(one(x),x)))
# @inline invlogit(x) = Base.FastMath.inv_fast(Base.FastMath.add_fast(one(x), exp(Base.FastMath.sub_fast(x))))
Expand All @@ -231,7 +231,7 @@ ldexp(x::Float16, q::Int) = Float16(ldexpk(Float32(x), q))
max_tanh(::Type{Float64}) = 19.06154746539849599509609553228539867418786340504817671278462587964799037885145
max_tanh(::Type{Float32}) = 9.010913339828708369989037671244720498805572920317272822795576296065428827978905f0

@inline function tanh_fast(x::AbstractSIMD{W,Float32}) where {W}
@inline function tanh_fast(x::Union{Float32,AbstractSIMD{<:Any,Float32}})
# stolen from https://github.com/FluxML/NNlib.jl/pull/345
# https://github.com/FluxML/NNlib.jl/blob/5dd04df4e95f9f9b70d6232fac546f3e98899fc2/src/activations.jl#L766-L773
x2 = abs2(x)
Expand All @@ -245,7 +245,7 @@ max_tanh(::Type{Float32}) = 9.01091333982870836998903767124472049880557292031727
d = muladd(d2, x2, 1.0f0)
ifelse(x2 < 66f0, @fastmath(x * (n / d)), sign(x))
end
@inline function tanh_fast(x::AbstractSIMD{W,Float64}) where {W}
@inline function tanh_fast(x::Union{Float64,AbstractSIMD{<:Any,Float64}})
exp2xm1 = expm1_fast(Base.FastMath.add_fast(x, x))
# Division is faster than approximate inversion in
# t = Base.FastMath.mul_fast(exp2xm1, Base.FastMath.inv_fast(Base.FastMath.add_fast(exp2xm1, typeof(x)(2))))
Expand All @@ -254,6 +254,12 @@ end
end
@inline tanh_fast(x::IntegerType) = tanh_fast(float(x))
@inline Base.FastMath.tanh_fast(x::AbstractSIMD) = tanh_fast(x)
@inline function Base.:(^)(
x::AbstractSIMD{W,<:Base.BitInteger},
y::AbstractSIMD{W,<:Base.BitInteger}
) where {W}
float(x) ^ y
end
# sigmoid_max(::Type{Float64}) = 36.42994775023704665301938332748370611415146834112402863375388447785857586583462
# sigmoid_max(::Type{Float32}) = 17.3286794841963099036462718631317335849086302638474573162299687307067828965093f0

Expand Down
6 changes: 3 additions & 3 deletions src/trig.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -191,7 +191,6 @@ end
c1 = -0.166666597127914428710938f0
return evalpoly(x, (c1, c2, c3, c4))
end

@inline function sin_fast(d::FloatType64)
T = eltype(d)
I = fpinttype(T)
Expand Down Expand Up @@ -234,7 +233,8 @@ end

return u
end

@inline sincos(x::IntegerType) = sincos(float(x))
@inline sincos_fast(x::IntegerType) = sincos_fast(float(x))
@inline function sin_fast(d::FloatType32)
T = eltype(d)
I = fpinttype(T)
Expand Down Expand Up @@ -904,7 +904,7 @@ Compute the inverse tangent of `x/y`, using the signs of both `x` and `y` to det
# return ifelse(isnan(y) | isnan(x), T(NaN), flipsign(r, x))
flipsign(r, x)
end

@inline atan_fast(a, b) = atan_fast(float(a), float(b))


"""
Expand Down
18 changes: 18 additions & 0 deletions test/testsetup.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -145,6 +145,15 @@ function test_vector(xfun, fun, ::Union{Val{W},SLEEFPirates.VectorizationBase.St
@test maximum(countulp.(t1, t2)) ≤ tol
@test maximum(countulp.(tu1, tu2)) ≤ tol
end
vui = round(SLEEFPirates.inttype(T), vu);
tu3 = tovector(xfun(vui));
tu4 = tovector(xfun(float(vui)));
@test maximum(countulp.(tu3, tu4)) ≤ tol
vxi = round(SLEEFPirates.inttype(T), vxes1);
tx3 = tovector(xfun(vxi));
tx4 = tovector(xfun(float(vxi)));
@test maximum(countulp.(tx3, tx4)) ≤ tol
nothing
end
vbig(x) = big.(x)
function test_vector(xfun, fun, ::Union{Val{W},SLEEFPirates.VectorizationBase.StaticInt{W}}, xf::NTuple{N,T}, xl::NTuple{N,T}, tol, broken::Bool) where {W,N,T}
Expand Down Expand Up @@ -190,6 +199,15 @@ function test_vector(xfun, fun, ::Union{Val{W},SLEEFPirates.VectorizationBase.St
else
@test_broken maximum(countulp.(tu1, tu2)) ≤ tol
end
vui = map(Base.Fix1(round,SLEEFPirates.inttype(T)), vu);
tu3 = tovector(xfun(vui...));
tu4 = tovector(xfun(map(float,vui)...));
@test maximum(countulp.(tu3, tu4)) ≤ tol
vxi = map(Base.Fix1(round,SLEEFPirates.inttype(T)), vxes1);
tx3 = tovector(xfun(vxi...));
tx4 = tovector(xfun(map(float,vxi)...));
@test maximum(countulp.(tx3, tx4)) ≤ tol
nothing
end
function test_function_acc(::Type{T}, xfun::F1, fun::F2, xx, tol, debug, tol_debug, broken) where {T,F1,F2}
rmax = 0.0
Expand Down

2 comments on commit 4a5c5ac

@chriselrod
Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@JuliaRegistrator register

@JuliaRegistrator
Copy link

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Registration pull request created: JuliaRegistries/General/69293

After the above pull request is merged, it is recommended that a tag is created on this repository for the registered package version.

This will be done automatically if the Julia TagBot GitHub Action is installed, or can be done manually through the github interface, or via:

git tag -a v0.6.36 -m "<description of version>" 4a5c5ac02d3be89b4c32134e93d535c3148516f8
git push origin v0.6.36

Please sign in to comment.